SCCN Home

SCCN Research Projects (1/31/2002)

The overall goal of the Swartz Center for Computational Neuroscience (SCCN) is to observe and model how functional activities in multiple brain areas interact dynamically to support human awareness, creativity and interaction.

Following are brief summaries of ongoing SCCN projects (as of Jan. 31, 2002).

Imaging event-related brain dynamics

Scott Makeig, Tzyy-Ping Jung

For forty years, research in human electrophysiology has been dominated by studies of averaged event-related potentials (ERPs) and fields (ERFs) recorded and averaged time locked to classes of experimental events. ERPs and ERFs provide unequivocal evidence of direct linkage between cognitive events and electrochemical brain events in a wide range of cognitive paradigms. However, collapsing a set of complex, multidimensional event-locked data epochs to a skeleton set of peak amplitude and latency measures of their average actually conceals rather than reveals many important types of event-related EEG and MEG brain dynamics. In fact, the standard 'ERP plus EEGnoise' model of evoked responses holds no mathematical privilege over any other linear decomposition of event-locked EEG data. This 'true-ERP' model would be a sufficient descriptor only if the EEG were a sum of noise processes with central tendency (e.g. Gaussian). However, the EEG has a low pass character, marked oscillatory features and a non-stationary temporal correlation structure. Changes in this structure induced by experimental events may produce the appearance of 'true' ERP (ERF) activity in the response average (Makeig et al., 2002).

That is, contrary to common assumption evoked responses may not be produced by brief synchronous neural activations in brain areas briefly engaged in successive stages of stimulus-related information processing. Rather, every feature of an evoked response may actually be produced by event-related changes in the autocorrelation and cross-correlation structure of ongoing EEG processes, each reflecting synchronous activity occurring continuously in one or more brain regions, or by more subtle perturbations in their dynamics. We will present evidence, derived by Independent Component Analysis (ICA), single-trial visualization, and time/frequency analysis, that most or all features of averaged responses following target and nontarget stimuli in a visual selective attention experiment may be reinterpreted as artifacts of averaging applied to event-related changes in the dynamics of ongoing EEG processes. These can be measured as event-related changes in the phase distribution of the EEG (or MEG) time locked to experimental events.

Historically, the focus on response averaging has contributed to an unnatural separation between the innately related sub-fields of ERP and EEG research, and has increased the conceptual separation between human electrophysiology and animal neurophysiology (supported as well by the neglect of local field potentials by neurophysiologists who record them). Currently, however, a new field is emerging that we propose to call cognitive event-related brain dynamics. Ever-increasing computational resources allow new generations of cognitive neuroscientists to study the dynamics of the scalp-recorded electromagnetic field using combinations of complex and sophisticated methods including source imaging, time/frequency analysis, event-related coherence, non-linear dynamics and independent component analysis. Applying these new techniques reveals that the EEG (and MEG) data contain a much unmined information about mechanisms of neural synchronization within and between brain areas. Further, this information appears to be convergent with new findings both in cellular neurophysiology and hemodynamic brain imaging. The new approach also makes feasible new experimental paradigms for studying the macroscopic brain dynamic substrates of a wide range of human cognitive processes.


Interrelationship between P300 and alpha rhythms in visual responses

Tzyy-Ping Jung, Erik Visser, Marissa Westerfield, Scott Makeig

Independent Component Analysis, performed on 600 single one-second target-centered EEG trials from a visual spatial selective attention task, extracted maximally independent components representing contributions from eye and muscle artifacts and from brain networks producing ongoing EEG and event-related potentials. The resulting components of 15 subjects were clustered into groups according to similarities between component scalp maps and activity spectra. For each subject and trial, activity of components in the cluster(s) accounting for most of the averaged P300 potential were back-projected onto and summed at each scalp channel. For every subject, in 10-30% of the trials the resulting P300-dominated data did not resemble the averaged P300-component activation. In these 'inconsistent-P3' trials, although the subject responded correctly to the target stimulus, the 'P300' appeared to be absent. Evoked activity time-locked to the stimulus (P1/N1) was quite similar in the inconsistent-P3 and consistent-P3 trial subsets. However, the post-stimulus phase distribution of alpha activity was more strongly realigned to the stimulus in the inconsistent-P3 trials. These results clearly demonstrate an interaction in single trials between event-related potentials (ERPs) and the ongoing EEG, an important fact that has been largely ignored by traditional ERP and EEG comunities.


Simultaneous recording and analysis of EEG and fMRI signals

Tzyy-Ping Jung, Jeng-Ren Duann, Marissa Westerfield and Scott Makeig

Electromagnetic and functional magnetic resonance brain imaging measures have complementary strengths and weaknesses in temporal and spatial resolution. First approaches to combining electrophysiological and hemodynamic brain data compared the scalp topography of event-related potential (ERP) averages against task-related fMRI activation level differences collected separately (George et al., 1995; Buckner et al., 1996; McCarthy et al., 1997). Shortcomings of this approach include difficulty in co-registration, and possible order effects in separately recorded data sets. More importantly, averaged ERP data reveal only a small portion of event-related brain dynamics (Makeig, 1993), and these may not replicate exact across sessions. We used a custom EEG amplifier with timeout circuits and shielded cabling (SA Instruments) to record single-trial event-related potentials (ERPs) at 22 scalp sites during EPI fMRI scanning from a healthy adult subject. A pilot subject participated in five five-minute bouts in which the subject performed the two-back continuous performance task (CPT) in 40-s time blocks alternating with 20-s blocks of (no-task) fixation. Independent Component Analysis (ICA) identified fMRI BOLD signal components with regions of activity (ROAs) in anterior and posterior cingulate cortex, respectively. In accord with recent reports (Raichle et al., 2001), BOLD signal levels of these midline components were were higher during the no-task blocks than during task performance. The time course of a third fMRI component was positively correlated with the experimental (task off-on) block design. Bilteral regions of activation (ROA) for this component were located in posterior parietal and motor cortex. After EEG-artifact rejection, changes in EEG power at frequencies above 10 Hz were strongly negatively correlated with the BOLD course, particularly in electrodes above the BOLD regions of cortical activation. We are currently collecting data to explore the robustness of these results.


From single-trial EEG to macroscopic brain dynamics

Arnaud Delorme, Scott Makeig

To assess how brain processes might interact and cooperate in the brain we applied new statistical techniques to single-trial EEG data from a fast go-nogo categorization task using natural images. The data we used were recorded by the team of Thorpe and Fabre-Thorpe. EEG recorded during the sessions showed an early difference between targets and non-target trials beginning at 150 ms after stimulus onset. The new Independent Component Analysis (ICA) method separates the different brain processes affected by the task, thus indexing changes in synchronization within and between cortical areas involved in rapid visual categorization. Our objectives are first to localize the different sources responsible for the observed event-related dynamics, and then to determine the dynamic relationship between these sources. We use time-frequency measures to assess the remaining temporal correlations between these sources and the time course of their interactions. We visualize the resulting pattern of activation and interaction using animations. These new decomposition and visualization techniques may prove critical for understanding macroscopic brain dynamics associated with high-level cognition.


Automatic EEG artifact rejection using high-order statistics and independent component analysis

Arnaud Delorme, Scott Makeig

While it is now generally accepted that independent component analysis (ICA) is a good tool for isolating both artifacts and cognitive related activations in EEG data (Jung et al., 1999, 2000, 2001), there is still little consensus about criteria for automatic rejection of artifactual components and of single trials dominated by artifact. Therefore, we developed five measures to detect artifactual trials: simple activity thresholding, linear trend detection, distribution of activity detection (using kurtosis), improbable event detection using joint-probability and frequency thresholding. We applied these to single trials using user-defined frequency windows. We then simulated artifacts of different amplitudes within actual EEG data and attempted to detect them automatically using combinations of the five measures. We assessed the performance of each measure for artifact detection. We also showed that preprocessing data using ICA improves detection of low amplitude artifacts on simulated data.


EEGLAB: A Matlab toolbox for advanced EEG data processing

Arnaud Delorme, Scott Makeig

We developed graphic user interface software called EEGLAB, running under MATLAB (The Mathworks, Inc.), to process single trial and averaged EEG data. The uses of the software range from basic EEG data processing (resampling, epoch extraction and averaging) and visualization (single-trial data visualization, 2-D and 3-D head plots) to more advanced data processing and visualization tools (ICA, time-frequency decomposition, advanced artifact rejection using spectral thresholding on multitaper spectra, etc.). The interface makes use of the ICA/EEG toolbox of Makeig et al (cnl.ucsd.edu/eeglab/). The EEGLAB software is organized in three layers of complexity for different types of users. At the top-level, 'push-button users' can deal with the graphic interface only, without extensive knowledge of Matlab syntax. At the intermediate level, user can make use the individual graphic commands write macros and automate processing of multiple data sets. More expert users can use the underlying standalone Matlab functions that perform the data processing directly. However, even expert users can take advantage of the top- and middle-level structure of the EEGLAB software to perform powerful fnd lexible EEG analysis. Thus, we believe that the EEGLAB software can benefit a wide range of researchers with different levels of expertise.


FMRLAB: A Matlab toolbox for fMRI data analysis

Jeng-Ren Duann, Tzyy-Ping Jung, and Scott Makeig

Analysis of functional magnetic resonance imaging (fMRI) brain data is a challenging enterprise, as the fMRI signals have varied, unpredictable time courses that represent the summation of signals from hemodynamic changes as a result of neural activities, from subject motion and machine artifacts, and from physiological cardiac, respiratory and other pulsations. The relative contribution and exact form of each of these components in a given session is largely unknown to the experimenter, suggesting a role for data-driven methods if the data have properties that are consistent with these models. The aims of the project are twofold: (1) To test applications of Independent Component Analysis (ICA) to concurrent fMRI and EEG data, first separately, by comparing the resulting fMRI source distributions with EEG source distributions derived from the ICA results using currently available EEG source localization approaches, and then jointly, by directly comparing the time courses of changes in fMRI activation and in the EEG frequency power spectrum. (2) To develop, test, document and distribute a software toolbox, based on the widely-used MATLAB signal processing environment, for carrying out the analyses we envision and visualizing the results. The MATLAB (The Mathworks, Inc.) tools we have developed allow researchers to decompose concurrently or separately recorded EEG and fMRI data into spatially and/or temporally independent components, and to evaluate their relationships to concurrent measures of task performance or other physiology and behavior. The analysis of electrophysiological and functional imaging data in both time and time/frequency domains will allow researchers to determine and evaluate relationships between ongoing or event- related changes in electrophysiology, hemodynamics and behavior and perform exploratory and hypothesis-driven analyses on human brain data from both basic and clinical studies of human brain and cognitive function.


Single-trial variability in event-related BOLD signals

Jeng-Ren Duann, Tzyy-Ping Jung, Scott Makeig

Most current analysis methods for functional magnetic resonance imaging (fMRI) data assume a priori knowledge of the time course of the hemodynamic response (HR) to experimental stimuli or events in brain areas of interest. In addition, they typically assume homogeneity of both the HR and the non-HR "noise" signals, both across brain regions and across similar experimental events. When HRs vary unpredictably, from area to area or from trial to trial, an alternative approach is needed. Here, we use infomax Independent Component Analysis (ICA) to detect and visualize variations in single-trial HRs in event-related fMRI data. Six subjects participated in four fMRI sessions each in which ten bursts of 8-Hz flickering-checkerboard stimulation were presented for 0.5-s (short) or 3-s (long) durations at 30-s intervals. Five axial slices were acquired by a Bruker 3T magnetic resonance imager at inter-scan intervals of 500 ms (TR). ICA decomposition of the resulting blood oxygenation level dependent BOLD data from each session produced an independent component active in primary visual cortex (V1) and, in several sessions, another active in medial temporal cortex (MT/V5). Visualizing sets of BOLD response epochs with novel BOLD-image plots demonstrated that component HRs varied substantially and often systematically across trials as well as across sessions, subjects, and brain areas. Contrary to expectation, in four of the six subjects the V1 component HR contained two positive peaks in response to short-stimulus bursts, while components with nearly identical regions of activity in long-stimulus sessions from the same subjects were associated with single-peaked HRs. Thus, ICA combined with BOLD-image visualization can reveal dramatic and unforeseen HR variations not apparent to researchers analyzing their data with event-related response averaging and fixed HR templates (Duann et al., 2002).


Awareness changes and brain dynamics during sleep onset

Luca A. Finelli, Scott Makeig, Kenneth B. Campbell (University of Ottawa)

The transition from wakefulness to sleep (nonREM stage 1), is characterized by alterations of attention level and sensory awareness. Waking, sleep, and the transition from wakefulness to sleep are defined by typical patterns of changes in the electroencephalogram (EEG), which is the most sensitive and widely used sleep/wake state indicator. To assess how brain processes might evolve with behavioral changes in attention and sensory awareness observed during the wake-to-sleep transition period, we are applying new statistical and signal processing methods to single-trial event-related EEG responses in an auditory "odd-ball" task. We are analyzing the EEG of 10 healthy subjects (6 female, 4 male, mean age 23.7 y) recorded from 29 derivations during repeated task presentation in sleep onset periods. Event-related brain dynamics have traditionally been quantified by simple measures of averages of single trials time-locked to classes of experimental events. ERP analysis of these data are reported by Cote et al. (2002). However our work has shown that event-related potentials (ERPs) are incomplete measures of the underlying EEG dynamics. In fact ERPs, in large part, are not created by event-related bursts of potential in brain areas involved in brief information processing of discrete stimulus event information, as conventionally believed, but often are produced instead by changes in the phase distribution of ongoing EEG activity. This can be separated by independent component analysis (ICA) into components representing epicenters of neural synchrony. These in turn can be clustered, across subjects, into characteristic classes. Preliminary results show that at least two classes of brain processes are common to subjects performing the auditory recognition task during sleep onset. We are exploring the complex dynamics of these classes of EEG processes in relationship to subject performance and awareness.


Independent component analysis of EEG signal dynamics

Joern Anemueller, Scott Makeig

Independent componenet analysis (ICA) has been used extensively to decompose non-invasively measured brain signals into components of independent brain activity. However, the analysis has largely been based on the assumption of stationarity of the underlying brain dynamics, presupposing spatially fixed sources with no internal dynamics. The current project aims at relaxing these two assumptions. In the first part of the project, the independent neural sources are modeled as dynamic activation patterns with inherent spatio-temporal dynamics. This model gives rise to algorithms of convolutive ICA, a technique that has previously been investigated in the area of speech signal processing. Existing convolutive algorithms are adapted to be applicable to high-dimensional data and are subsequently employed to analyze human brain EEG signals. The second part of the project aims at relaxing the assumption of spatially fixed sources. Building on work recently begun by Makeig, Enghoff et al. (2000), a time-varying mixing system is assumed to model the source superposition. Algorithms for non-stationary ICA are developed and applied to EEG signal recordings. The results are interpreted in particular with respect to the changing spatial structure of the sources.